Академический Документы
Профессиональный Документы
Культура Документы
Polynomial Models
Author: Michael Rackl
Ostbayerische Technische Hochschule Regensburg
Mechanical Engineering Department
31st May 2015
E-Mail: rackl@fml.mw.tum.de
This document provides the theoretical background for the Scilab curve
fitting scripts for Ogden, Yeoh and Polynomial models, as presented at ScilabTEC conference 2015.
Please cite [Rackl (2015)], if you find these information useful.
Contents
1. Hyperelastic Material Models
1.1. Ogden Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.2. Polynomial Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.3. Yeoh Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1
1
1
1
2
6
6
8
4. Discussion
11
A. Appendix
A.1. Derivation of Stress-Strain Relationships . . . . . . . . . . . . . . . . . .
A.1.1. Yeoh Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
12
12
12
13
13
14
n
X
p p
1 + 2 p + 3 p 3
p=1 p
(1)
n
X
i
Cij I1 3 I2 3
i=0,j=0
j
, where C00 = 0.
(2)
This model is also called the generalized Rivlin model (Chang et al. 1991, Hartmann &
Neff 2003, Laksari et al. 2012).
A specific case of Equation 2 was developed earlier by Mooney (1940) and Rivlin
(1948b). It is referred to as the Mooney-Rivlin model and can be derived from Equation 2
by setting n = 1, C01 = C2 , C11 = 0 and C10 = C1 , which yields Equation 3 for the strain
energy W .
W = C1 I1 3 + C2 I2 3
(3)
n
X
i=1
i
Ci I1 3 .
(4)
model to represent the real behavior of the material, more than one loading mode should
be considered. In addition to that, the selected loading modes shall be similar to the
desired loading case of the application (Meier et al. 2003).
Curve fitting for hyperelastic material models divides into four subtasks. First of
all, the stress-strain curves need to be revised and adapted, where required. Second, a
constitutive equation (material model) has to be chosen, and third, an error criterion for
the goodness of fit has to be defined. The final step is to qualitatively and quantitatively
compare the resulting curve to the measured data.
The curve fitting process is similar for the Yeoh and Ogden hyperelastic models. The
stress-strain relationships for both models are derived in the appendix (subsection A.1).
2
2
The strain invariants, expressed in terms of the three stretch ratios 1 , 2 and 3 , are
(Treloar 1973)
I1 =21 + 22 + 33 ,
I2 =21 22 + 22 23 + 23 21
I3 =21 22 23 .
(6)
(7)
(8)
and
The stretch ratios i represent the deformation of a differential cubic volume element
along the principle axes of a Cartesian coordinate system. They are defined as the ratio
of the deformed length li to the undeformed length Li (Equation 9). The stretch ratio
equals 1 in undeformed state.
i =
li
Li
i [1, 2, 3]
(9)
Note that given the assumption of incompressibility of the material, the third strain
invariant I3 yields
I3 = 21 22 23 = 1,
(10)
(11)
(12)
2
Figure 1
A cubic differential volume element with
tensile stress and Cartesian coordinate
system.
tensile stress , deformations in 2 and 3 are equal (Treloar 1973). The corresponding
mathematical expressions are
1 =
2 =3 .
and
(13)
(14)
2 = 3 = 2 .
(15)
Equation 13 and Equation 15 can now replace 1 and 2 in Equation 11 and Equation 12. Resulting from this, the two strain invariants for an incompressible material in
and
(16)
(17)
The actual relation between engineering stress and stretch for an incompressible material under tension/compression is (Rivlin 1948b, 1956)
e = 2 2
1 W
W
+
.
I1
I2
(18)
Where:
e : Engineering stress
W : Strain energy
Stretch, parallel to e
Strain invariants
:
I1 , I2 :
Inserting Equation 5 into Equation 18, followed by calculation and simplification, gives
"
+ C11 I2 3 + I1 3
1
(19)
Finally, using Equation 16 and Equation 17, the two strain invariants I1 and I2 can
be eliminated. After simplification, the engineering stress-stretch relationship for
tension and compression is
"
+2C02 2 +
3 + 3C11 1
(20)
x
.
h0
(21)
Where h0 equals the thickness along axis 2 and x is the displacement along axis 1.
Let be the stretch in axis 1. The surfaces parallel to the 1-3 plane move parallel
against each other, along axis 1. Therefore the gap distance does not change. The
principal stretches 1 and 2 are (Rivlin 1948b, Treloar 1973)
1 =
2 = 1.
and
(22)
(23)
h0
3
Figure 2
A cubic differential volume element under shear stress . The amount of shear strain
is defined as the quotient of the displacement along axis 1 and the constant height h0 .
Again, assuming incompressibility, Equation 10 applies and 3 yields
3 = 1 .
(24)
The first and second strain invariant are the same for simple shear and can be expressed
in terms of the amount of shear strain (Rivlin 1948b). They are
I1 = I2 = 3 + 2 .
(25)
W
W
.
+
=2
I1
I2
(26)
(27)
which, in combination with Equation 25, yields the stress-strain relationship for
simple shear in Equation 28.
(28)
Curve Fitting aims to fit one or more parameters of a model equation , Parameters
, Parameters = M
(29)
Where , Parameters is the model function, which depends on stretch and parameters,
(30)
The parameters (material constants) Cij are considered constant and have to be determined through regression analysis. Furthermore, in the compression/tension model
equation, they only appear linearly. As the number of data pairs within the measured
stress-stretch curve greatly exceeds the number of parameters within the model equation, the problem definition is to solve an overdetermined system of linear equations
(Hartmann 2001) in such a way that a satisfactory goodness of fit is achieved.
n
X
i=1
With:
ls : Least square error,
i: Number of measured data pairs,
i : Measured stretch value,
!2
minimum
(31)
M i : Measured stress at i ,
i , C10 , C01 , C20 , C02 , C11 : Computed stress value of the model function at i .
In ANSYS1 , Equation 31 is called unnormalized least squares fit ANSYS Inc. (August
2002).
Since Equation 31 is biased towards higher stress values, a weighted error criterion is
more useful in many cases. Equation 32 accounts equally for every stress value and is
called normalized least square fit ANSYS Inc. (August 2002).
norm =
n
X
i=1
M i
2
minimum
(32)
Note that Equation 32 would lead to a division by zero if M i = 0. For this case
an exception has to be added. In this study, Equation 33 was chosen. It is the same
formulation as in the least square error in Equation 31. For small values around zero,
the normalized and the unnormalized least square criterion yield similar results.
!
2
for M i = 0
(33)
The software ANSYS offers a curve fitting module for hyperelastic material models.2
However, despite supporting higher order constitutive equations for input of material
constants, not all supported material models can be fitted up to these orders. Hence,
the Equations 20 and 28 and both error criteria3 were implemented in a Scilab4 script in
order to be able to fit higher order constitutive equations. The minimization algorithm
used in Scilab is Nelder-Mead (Nelder & Mead 1965).
Since two curves, one for tension/compression and one for simple shear, have to be
fitted in this study, the error used for the minimization process has to consist of the
respective errors of both curves. The Scilab script treats the error for both loading
scenarios equally (Equation 34).
min = tens/comp + shear
(34)
With:
min : Error used in the minimization process,
tens/comp : Error from combined tension and compression curve fitting,
shear : Error from simple shear curve fitting.
There exists a global analytic solution for the parameter set in the linear least square
problem of the compression/tension curve. However, there is no analytic solution for the
nonlinear least square problem that occurs when trying to fit the curve for simple shear
(Zielesny 2011, p. 41-42). The method applied in this study is suitable for both linear and
nonlinear problem types, and is easily adaptable for different constitutive hyperelastic
equations.
As the Ogden models lead to a nonlinear optimization problem which requires an
iterative solution, 1000 iterations were carried out for each model in ANSYS, as well as
in the Scilab script.
n1
R2 1
np1
n
P
Yi Yi
(35)
2
R2 = 1 i=1
2
n
P
Yi Yi
(36)
i=1
Where:
R2 : Adjusted coefficient of determination,
n: Number of data pairs,
p: Number of parameters in the model function,
R2 : Coefficient of determination,
Yi : Measured value,
Yi : Value from the model function,
Yi : Average of the measured values.
Table 1
Ogden models used in curve fitting and results for the material parameters in ANSYS
and using the Scilab script ([i] = MPa, [i ] = 1).
Order error
number of
criterion
parameters
ANSYS
Scilab script
norm.
1 = 0.1244
1 = 1.316
2 = 0.6225
2 = 1.299
1 = 4.290
1 = 0.1370
2 = 7.484 103
2 = 4.346
3rd
abs.
1 = 0.1770
1 = 1.055
2 = 0.2338
2 = 1.041
3 = 0.3251
3 = 0.9889
1 = 28.01 103
1 = 3.525
2 = 4.385
2 = 0.7777
3 = 7.270
3 = 0.5449
4th
abs.
not available
1 = 58.90 103
1 = 2.800
2 = 3.552
2 = 0.3355
3 = 4.380
3 = 79.13 103
4 = 0.4110
4 = 0.9359
nd
Table 2
Polynomial models used in curve fitting and results for the material parameters in ANSYS and using the Scilab script. The model with three parameters is equivalent to a
three parameter Mooney-Rivlin model. ([Cij ] = MPa).
Order error
number of
criterion
parameters
ANSYS
Scilab script
1st
norm.
abs.
2nd
abs.
C10
C01
C20
C11
C02
= 88.41 103
= 0.2197
= 0.2122
= 0.2909
= 78.22 103
C10
C01
C20
C11
C02
= 88.41 103
= 0.2197
= 0.2122
= 0.2909
= 78.22 103
Table 3
Yeoh models used in curve fitting and results for the material parameters in ANSYS
and using the Scilab script ([Ci0 ] = MPa).
Order error
number of
criterion
parameters
ANSYS
Scilab script
rd
norm.
C10 = 0.1595
C20 = 6.930 103
C30 = 0.4030 103
C10 = 0.1595
C20 = 6.930 103
C30 = 0.4030 103
4rd
norm.
not available
C10
C20
C30
C40
= 0.1593
= 6.220 103
= 0.2210 103
= 11.90 106
5rd
norm.
not available
C10
C20
C30
C40
C50
= 0.1573
= 2.391 103
= 3.907 103
= 0.6600 103
= 32.60 106
10
4. Discussion
Scilab scripts were created in order to be able to fit experimental stress-strain data
to higher order hyperelastic material models than ANSYS allows for. For lower order
models, the scripts were compared against the parameter sets provided by ANSYS.
The Scilab parameter sets for the polynomial and the Yeoh model are consistent
with those from ANSYS (see Table 3 and Table 2). This indicates that the stressstrain equations were derived and implemented correctly, as well as that the regression
algorithm and error criteria yield the same results.
However, discrepancies were discovered when fitting the Ogden material models (see
Table 1). When examining these differences, a 1st order Ogden model was fitted with
absolute and normalized error criteria.5 For both criteria, the Scilab script yields
similar results to what ANSYS provides. Table 4 shows a comparison of the four
parameter sets.
Table 4
Ogden models material parameters, fitted in ANSYS and using the Scilab script ([i] =
MPa, [i] = 1).
Order error
1
st
1st
number of
criterion
parameters
ANSYS
Scilab script
abs.
1 = 0.7055
1 = 1.0546
1 = 0.7208
1 = 1.0345
norm.
1 = 0.5571
1 = 1.1609
1 = 0.6356
1 = 1.023
These data suggest that the discrepancies are not due to false implementation of the
stress-strain relationships of the Ogden models, but because of other differences, such as
the optimization algorithm used and the nonlinear nature of the problem. In particular,
the latter combination and the existence of more than one minimum makes it difficult
to effectively compare both curve fitting algorithms for the Ogden model. Even using
identical start values for the iteration process may lead to completely different material
parameter sets (Zielesny 2011, p. 146). This effect increases with higher order models.
The curve fitting algorithms presented in this study allow fitting of compression/tension
and simple shear test data to any order Ogden models, any order Yeoh models and 2nd ,
as well as 1st , order polynomial models. As a special case of the polynomial model, a
three parameter Mooney-Rivlin model may be fitted, too. Contrary to the curve fitting
module in ANSYS, the Scilab scripts allow for biasing of one of the two loading cases
(see Equation 34) and increased adjustability of optimization parameters.
In addition to that, auxiliary conditions for the material constants could be considered
by extending the source code. However, this was not taken into account within this study
in order to be able to compare the results to those from ANSYS. Furthermore, as the
5
Both of these models were no longer considered after the preselection of material parameter sets, due
to a bad fit.
11
A. Appendix
This appendix contains further information on the derivation of stress-strain relationships for the Ogden and Yeoh material models, as well as the curve fitting results, which
were not considered for use within this study.
n
X
i
Ci I1 3 .
i=1
(37)
n
X
2 Ci i 2 I1 3
i=1
i1
(38)
e =
n
X
i=1
2 Ci i 2 2 + 21 3
12
i1
(39)
n
X
2 Ci i I1 3
i=1
i1
(40)
If one combines Equation 40 with Equation 25, this yields the stress-strain relation for
simple shear of a nth order Yeoh model, which is
n
X
2 Ci i 2(i1) .
(41)
i=1
n
X
p p
1 + 2 p + 3 p 3
p=1 p
(42)
W =
n
X
1
p p
+ 2 2 p 3 .
p=1 p
(43)
According to Doghri (2000), as cited in Stommel et al. (2012, p. 77), the stress-strain
relation for compression and tension can be obtained by deriving the strain energy with
respect to the stretch (Equation 44).
(44)
n
X
p=1
p 1
( 12 p +1)
(45)
W =
n
X
p=1
p i + i 2 .
(46)
13
Again, Equation 44 applies and finally yields to Equation 47 as the simple shear
stress-strain relationship for a nth order Ogden model. (Ogden 1972)
n
X
p=1
p
+ 1
(47)
The stretch in simple shear was not measured during the material tests. Ogden (1972)
showed that the simple shear strain can be expressed in terms of the stretch as follows:
1 = .
(48)
Using Equation 48, the stretch cannot be directly calculated by solving the equation for
. However, Equation 48 may be harnessed in order to iteratively compute the amount
of stretch from a given amount of shear. The corresponding function code in Scilab is:
// compute stretch from given amount of shear
// Input: a simple shear strain value
<gamma>
// Output: a stretch value
<lambda>
function out = ConvToStretch(gamma)
lambda = 1;
while abs((gamma+lambda^(-1)-lambda))>1e-6
lambda = gamma + lambda^(-1);
end
out = lambda;
endfunction
References
ANSYS Inc. (August 2002), ANSYS material modeling: Hyperelastic material characterization. , last checked 20.08.2013.
URL:http://ansys.net/ansys/papers/nonlinear/hyper_elasticcity_
curvefitting.pdf
ANSYS Inc. (n.d.), ANSYS Mechanical User Manual.
Brown, R. (2006), Physical testing of rubber, 4 edn, Springer, New York.
Chang, T. Y. P., Saleeb, A. F. & Li, G. (1991), Large strain analysis of rubber-like materials based on a perturbed Lagrangian variational principle, Computational Mechanics
8(4), 221233.
Doghri, I. (2000), Mechanics of Deformable Solids: Linear, Nonlinear, Analytical and
Computational Aspects, Springer Berlin Heidelberg, Berlin and Heidelberg.
Fahrmeir, L., Kneib, T. & Lang, S. (2009), Regression: Modelle, Methoden und Anwendungen, Statistik und ihre Anwendungen, 2 edn, Springer, Berlin and Heidelberg.
14
15
Rivlin, R. S. & Saunders, D. W. (1951), Large Elastic Deformations of Isotropic Materials. VII. Experiments on the Deformation of Rubber, Philosophical Transactions of
the Royal Society A: Mathematical, Physical and Engineering Sciences 243(865), 251
288.
Schwarzl, F. R. (1990), Polymermechanik: Struktur und mechanisches Verhalten von
Polymeren, Springer, Berlin.
Selvadurai, A. (2006), Deflections of a rubber membrane, Journal of the Mechanics and
Physics of Solids 54(6), 10931119.
Stommel, M., Stojek, M. & Korte, W. (2012), FEM zur Berechnung von Kunststoff- und
Elastomerbauteilen, Hanser, Mnchen.
Treloar, L. R. G. (1973), The elasticity and related properties of rubbers, Reports on
Progress in Physics 36(7), 755826.
Yeoh, O. H. (1993), Some Forms of the Strain Energy Function for Rubber, Rubber
Chemistry and Technology 66(5), 754771.
Zielesny, A. (2011), From curve fitting to machine learning: An illustrative guide to
scientific data analysis and computational intelligence, Vol. 18 of Intelligent systems
reference library, Springer, Berlin.
16